coulomb/intelligence-radar
5
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions
intelligence-radar/digests/daily/2026-09-18.md

15 KiB
Raw Blame History

Intelligence Radar Daily 2026 09 18

Intelligence Radar — DAILY BRIEF
18 September 2026 (Europe/Berlin)

Frontier weather: dense, post-burst consolidation. Early-September dual launches of Claude Fable 5.1 and GPT-6 Astra continue to define the available ceiling on agentic coding and knowledge work; open-weight efficiency pressure from DeepSeek V4.1 Flash is immediate and measurable.

1. Frontier & markets

Confirmed releases cluster tightly. Anthropic shipped Claude Fable 5.1 (and invitation-only Mythos 5.1) on 1 September: same $10 / $50 per MTok headline pricing as Fable 5, but cache reads cut to $0.25 / MTok (2.5 % of input). Context remains 1 M tokens, adaptive thinking always on, knowledge cutoff June 2026. System card and model card confirm stronger long-horizon agentic coding and research capabilities; Mythos variant relaxes dual-use classifiers for verified cybersecurity and life-sciences users.

OpenAI followed on 3–4 September with GPT-6 Astra (API id gpt-6-astra). Standard pricing matches Fable at $10 input / $50 output (cache $1 / $12.50); long-context (>272 k input) doubles most rates. Context 1.05 M tokens. Roll-out reached Plus/Pro/Business/Enterprise and Azure/Bedrock within days; Fast mode at 2× price for ~2× speed. OpenAI rates it Critical under its Preparedness Framework for cybersecurity.

DeepSeek released V4.1 Flash (MIT open weights) on 10 September: 552 B MoE backbone, Causal Encoder–Decoder architecture activating ~8 B parameters on prefill / 16 B on decode, native vision, 1 M context, KV cache compressed to ~890 bytes/token. Peak pricing ~$0.30 / $1.20 (off-peak half); API id deepseek-flash. DeepSeek reports it ahead of its own prior V4-Pro on multiple agentic metrics and is phasing the older Pro routing.

Meta Muse Spark 1.3 (2 September) and Google Gemini 3.8 Flash / Live variants filled the same window. Artificial Analysis Intelligence Index v4.3 (updated early September, incorporating Terminal-Bench 4.0 and AutomationBench-AA) places both Fable 5.1 (max + fallback) and GPT-6 Astra (max) at 53, with Astra delivering the same score at materially lower cost-per-task. Open-weight leaders remain GLM-5.x and Qwen3.8 variants in the low-40s. Never mix Index versions; v4.3 is the current public board.

Availability remains capacity-constrained at the absolute frontier; region blocks and export-control residual effects from earlier 2026 still shape certain Mythos-class access. Cache-price cuts (Anthropic) and extreme compression (DeepSeek) are the clearest near-term economic events.

2. Science of intelligence

Measurement continues to migrate away from saturated static exams toward interactive and agentic harnesses. ARC-AGI-3 (interactive reasoning) saw GPT-6 Astra report 99.9 % under the Provider Adapter at high effort — a large jump from prior GPT-5.6-class scores — while the ARC Prize 2026 Milestone #2 window remains open until 30 September. Chollet’s framing (skill-acquisition efficiency under novelty) remains the explicit reference for these boards; saturated ARC-AGI-1 numbers are no longer treated as rankers.

Primary theoretical work remains sparse this week. Tenenbaum’s Bayesian cognitive-modeling program (Daedalus 2026) continues to explore integration of LLMs with probabilistic world models. Ganguli’s call for a unified physics–neuroscience–AI science of intelligence (same volume) emphasizes emergent collective dynamics and the still-large efficiency gap between biological and artificial systems. New axiomatic proposals (SANC(E₃), quantitative intelligence-density measures) appear on arXiv but have not yet shifted laboratory evaluation practice. No major primary publications from Amodei, Hassabis, Clark, or Epoch authors landed in the last 48 h that alter measurement doctrine.

3. Engineering

Agent harnesses dominate. Terminal-Bench 2.1 / 4.0 and OSWorld-Verified remain the practical signal for long-horizon tool use; Cognition’s SWE-2 leads certain Terminal-Bench snapshots, while Qwen3.8 Max and Claude Fable variants trade the top of OSWorld-Verified. DeepSeek’s KV-cache compression and Anthropic’s cache-price cut are the clearest inference-cost events. Nvidia continues power-efficiency messaging around Vera Rubin / DSX MaxLPS (higher tokens per megawatt, flexible load programs). Custom silicon roadmaps (OpenAI–Broadcom Jalapeño, Meta MTIA generations) remain multi-quarter away from broad availability; current economics are still dominated by software-level cache and sparsity wins.

Energy and capacity remain first-class constraints: token throughput per megawatt is the metric hyperscalers now track publicly.

4. Political / legal / social

EU AI Act GPAI obligations and AI Office enforcement powers are live; transparency rules (labelling, chatbot disclosure) apply. The Digital Omnibus deferred certain high-risk system deadlines. US federal posture remains light-touch / voluntary (EO framework); California frontier-transparency rules continue in force. G20 discussions this month reflected US preference for non-technology-specific principles versus EU product-safety framing. No new binding global licensing regime emerged. Cybersecurity capability thresholds (OpenAI Critical rating for Astra; Anthropic dual-use classifiers) are now operational access controls rather than purely internal policy.

5. Watchlist hits (primary)

6. Ledger actions

  • Frontier label: retain Claude Fable 5.1 and GPT-6 Astra (AA Intelligence Index v4.3 = 53; task-dependent on Terminal-Bench / OSWorld / ARC-AGI-3). No single best model without a specified board and harness.
  • Efficient label: elevate DeepSeek V4.1 Flash on cost-per-task and open-weight availability; monitor GLM / Qwen3.8 Max for agentic boards.
  • Budget label: unchanged — Flash-tier and open-weight mid-range remain the practical floor.
  • Five-benchmark atlas (AA Index v4.3, Terminal-Bench 4.0, OSWorld-Verified, ARC-AGI-3, SWE-bench / coding-agent suite) requires no rebase this cycle; cache economics and KV compression are the variables that would move Efficient/Budget thresholds first.

Ask Intelligence Radar a task for a Task Router ticket (5 benchmarks → Frontier / Efficient / Budget + optional decomposition).

Intelligence Radar — Daily Delta, 18 Sep 2026

Net assessment: there is no verified new general-purpose flagship model or broad API-price change since the previous scan. The material change is more structural: Anthropic has published the clearest quantitative picture yet of AI helping to build its own successors, OpenAI has shown how much a domain-configured system can outperform its underlying frontier model, and Anthropic has produced a concrete example of AI materially lowering the cost of scientific-model engineering.

1. Anthropic publishes a metrology for “AI building AI”

On September 17, Anthropic published three internal measures intended to make the pace of frontier development externally legible: the share of AI R&D performed by AI, the quality of oversight over internal agents, and the allocation of compute between capability and safety work. (Anthropic)

Its most important result is the new Anthropic R&D Automation Index. As of August 2026, Claude did not operate fully autonomously on any measured category, but it “led” 26% of Anthropic’s AI R&D work—meaning it could perform most of a task end-to-end from a high-level prompt with human supervision—and more than 90% of R&D work was at least at the “AI collaborates” level. The underlying chart says the “leads” share had risen from under 1% in February. (Anthropic)

Anthropic also reports roughly 30,000 concurrent research/engineering agents on its most-used internal platform. All their actions pass through online monitoring; in August, the company analyzed more than a billion agent decisions and says 0.002%—about 1 in 47,000—were blocked. Offline monitoring escalates roughly 50 high-priority cases per week to humans. (Anthropic)

Its compute snapshot is also novel: during one measured week, Anthropic classified about 6% of AI-R&D compute as safety work, rising to about 12% within AI-driven AI R&D. Anthropic explicitly warns that these shares are imperfect and methodology-dependent. (Anthropic)

Radar impact: yesterday’s recursive_self_improvement watch item can now be sharpened considerably. The evidence supports:

AI-assisted recursive development is already substantial and rapidly increasing, but fully autonomous successor development has not been observed in Anthropic’s own measurement.

So I would classify:

  • AI_accelerates_AI_R&D → observed / strong
  • AI_leads_substantial_fraction_of_frontier_R&D → observed / provider-measured
  • fully_autonomous_recursive_self_improvement → not observed
  • recursive_self_improvement_is_imminent → contested/watch

More importantly, Intelligence Radar should add a new Development Process Metrics family alongside capability benchmarks:

R&D automation level · agent count · oversight coverage · review latency · escalation/block rate · safety-compute share

These are not conventional model benchmarks; they measure the production system that creates the next generation of models.

Anthropic — Measurements for understanding the pace of AI development inside frontier labs

2. Astra for Law demonstrates that the execution system can matter almost as much as the base model

OpenAI introduced Astra for Law on September 17. Crucially, this is not a new foundation model. It combines GPT‑6 Astra with legal-specific instructions and a dedicated search index covering U.S. case law, statutes, regulations, court rules and administrative decisions across more than 230 million URLs. (OpenAI)

On 200 private validation questions from Vals AI’s Legal Research Bench, OpenAI reports that the complete Astra-for-Law configuration passed the benchmark’s overall correctness check on 54.0% of questions, versus 38.7% for ordinary GPT‑6 Astra with web search at the same highest reasoning setting—a 40% relative improvement. It also found 24% more reference cases on case-law questions. (OpenAI)

Availability is initially restricted to selected U.S. law firms through Trusted Access; API access is planned but not yet generally available. (OpenAI)

Radar impact: this strongly validates our decision not to rank only “models.”

For legal research, the measured entity is:

GPT‑6 Astra + legal instructions + specialized corpus/index + retrieval + access policy

not simply GPT‑6 Astra.

I would make one taxonomy refinement:

Execution Profile → Domain Profile

A Domain Profile is an execution profile whose capability comes materially from specialist retrieval, tools, instructions, or workflow constraints.

The result also creates a benchmark warning: generic frontier-model scores can substantially underpredict task performance once a high-quality domain environment is added. The five-benchmark selector should therefore prefer evaluations whose environment resembles the proposed deployment.

Astra for Law does not change the global Frontier ranking, but it likely changes the Available Frontier for U.S. legal research for users with access.

OpenAI — Introducing Astra for Law

3. Claude materially reduces the computational cost of biomolecular AI

Anthropic also published a substantial science/engineering result on September 17. Working with its Claude Science setup, Claude optimized more than 30 open-source biomolecular deep-learning models in under four weeks, achieving roughly 4× average speed-up with minimal precision loss and nearly 2× with identical outputs. (Anthropic)

For protein-structure systems, Anthropic reports custom kernels outperforming the existing field standard by roughly 2.7–2.9× for triangle attention and 1.7–3.2× for triangle multiplication. Claude also created a low-memory mode enabling accurate prediction of biomolecular systems above 10,000 tokens on a single GPU node. (Anthropic)

The economics are particularly relevant. Anthropic’s earlier agentic protein-design setup allowed up to roughly 2,500 H100 GPU-hours per target. With the new optimized tooling, a single Claude model using one H200 for 24 hours reportedly achieved comparable in-silico design scores using roughly 100× fewer GPU-hours, with about $150 total GPU-plus-model spend. This is computational evidence only; wet-lab validation is still the important downstream test. (Anthropic)

Radar impact: this exposes another benchmark gap. Scientific capability should not be represented only by answering scientific questions or writing research code. Add constructs such as:

scientific_toolchain_engineering domain_model_optimization compute_efficiency_improvement AI_improving_other_AI_systems

This also reinforces the previous finding about AI-assisted AI R&D: a general-purpose frontier model can improve specialized learned systems, not merely write glue code around them.

Anthropic — How Claude is uplifting biomolecular modeling

Recommended Intelligence Radar state changes

I would make three concrete changes today: add Development Process Metrics as a new measurement family; introduce Domain Profile as a named subtype of Execution Profile; and expand scientific capability measurement from reasoning/programming into optimization of scientific AI systems and their resource economics.

There is no reason from today’s evidence to alter the general Frontier / Topnotch / Efficient / Budget baseline. The more consequential delta is that the unit of analysis keeps widening:

model → execution profile → domain system → development organization.

That increasingly suggests Intelligence Radar should measure not only what an AI can do, but also how much of the machinery that produces the next AI is itself being operated by AI.